Affiliation:
1. School of Electronic and Electrical Engineering, Hongik University, Seoul 04066, Republic of Korea
Abstract
This study presents a charge-domain SRAM-based in-memory computing (IMC) architecture. The multiply-and-accumulate (MAC) operation in the IMC structure is divided into current- and charge-domain methods. Current-domain IMC has high-power consumption and poor linearity. Charge-domain IMC has reduced variability compared with current-domain IMCs, achieving higher linearity and enabling energy-efficient operation with fewer dynamic current paths. The proposed IMC structure uses a 9T1C bitcell considering the trade-off between the bitcell area and the threshold voltage drop by an NMOS access transistor. We propose an energy-efficient summation mechanism for 4-bit weight rows to perform energy-efficient MAC operations. The generated MAC value is finally returned as a digital value through an analog-to-digital converter (ADC), whose performance is one of the critical components in the overall system. The proposed flash-successive approximation register (SAR) ADC is designed by combining the advantages of flash ADC and SAR ADC and outputs digital values at approximately half the cycle of SAR ADC. The proposed charge-domain IMC is designed and simulated in a 65 nm CMOS process. It achieves 102.4 GOPS throughput and 33.6 TOPS/W energy efficiency at array size of 1 Kb.
Funder
Ministry of Science and ICT
Reference33 articles.
1. 8T SRAM Cell as a Multibit Dot-Product Engine for Beyond Von Neumann Computing;Jaiswal;IEEE Trans. Very Large Scale Integr. (VLSI) Syst.,2019
2. 10T SRAM Computing-in-Memory Macros for Binary and Multibit MAC Operation of DNN Edge Processors;Nguyen;IEEE Access,2021
3. Si, X., Tu, Y.-N., Huang, W.-H., Su, J.-W., Lu, P.-J., Wang, J.-H., Liu, T.-W., Wu, S.-Y., Liu, R., and Chou, Y.-C. (2020, January 16–20). 15.5 A 28nm 64Kb 6T SRAM Computing-in-Memory Macro with 8b MAC Operation for AI Edge Chips. Proceedings of the 2020 IEEE International Solid-State Circuits Conference—(ISSCC), San Francisco, CA, USA.
4. A 4+ 2T SRAM for Searching and In-Memory Computing with 0.3-V VDDmin;Dong;IEEE J. Solid-State Circuits,2017
5. A Drift-Resilient Hardware Implementation of Neural Accelerators Based on Phase Change Memory Devices;Bianchi;IEEE Trans. Electron Devices,2021
Cited by
3 articles.
订阅此论文施引文献
订阅此论文施引文献,注册后可以免费订阅5篇论文的施引文献,订阅后可以查看论文全部施引文献